“`html
Nvidia is reportedly developing a new AI chip tailored for the Chinese market, potentially exceeding the performance of its existing H20 offering, Reuters reports. Sources familiar with the matter indicate the semiconductor giant is leveraging its cutting-edge Blackwell architecture for this endeavor.
The move comes amid ongoing US-China trade tensions and stringent export controls on advanced AI technology. Last week, speculation arose after former President Donald Trump hinted at the possibility of Nvidia selling more advanced chips in China, although regulatory approvals remain a significant hurdle given Washington’s hesitation to grant Beijing unfettered access to US AI prowess.
Internally dubbed the B30A, the new chip is expected to adopt a single-die design, consolidating all core components onto a single silicon substrate. This contrasts with Nvidia’s flagship B300, which utilizes a dual-die configuration. Insiders suggest this design choice would position the B30A’s performance at roughly half that of the B300, but still surpassing the H20 currently available in China.
Mirroring the H20, the B30A will incorporate high-bandwidth memory and Nvidia’s NVLink technology, facilitating rapid data transfer between processors. While final specifications are still being finalized, Nvidia aims to dispatch samples to Chinese clients for evaluation as early as next month. This aggressive timeline underscores the company’s intent to maintain its presence in the crucial Chinese market.
In an official statement addressing its China strategy, Nvidia emphasized its commitment to compliance. “We evaluate a variety of products for our roadmap, so that we can be prepared to compete to the extent that governments allow,” the statement read. “Everything we offer is with the full approval of the applicable authorities and designed solely for beneficial commercial use.”
US Commerce Secretary Howard Lutnick told CNBC that Nvidia CEO Jensen Huang is actively advocating for approval of the new chip. “Of course he would like to sell a new chip to China,” Lutnick stated, noting Huang’s frequent direct appeals to the president. “The president listens to our great technology companies, and he’ll decide how he wants to play it. But the fact Jensen is pitching a new chip shouldn’t surprise anybody.” Lutnick’s comments highlight the ongoing behind-the-scenes lobbying efforts surrounding semiconductor exports to China.
US-China trade tensions grow over Nvidia chips
The degree of China’s access to advanced AI chips has become a key point of contention in US-China relations. China represents a significant portion of Nvidia’s revenue, accounting for 13% of its sales last year, making the market a critical battleground for the company.
Nvidia’s H20 sales in China only resumed in July, following a temporary halt by US regulators in April. The H20 was specifically engineered in 2023 to comply with export regulations restricting chip sales to China, showcasing the tightrope Nvidia walks in attempting to retain market share while adhering to US policy.
Former President Trump’s recent suggestion that Nvidia might be permitted to sell a less potent version of its next-generation chip in China adds another layer of complexity. His proposal, potentially tied to a broader agreement, included Nvidia and rival AMD ceding 15% of revenue from certain chip sales to China to the US government. Trump also dismissed the H20 as “obsolete,” suggesting a new chip could offer a performance level “30% to 50% off” the top-tier model.
Bipartisan voices in Washington argue that even watered-down AI chips could provide China with a competitive advantage in critical sectors. Nvidia and other chipmakers counter that restricting sales to China will simply push Chinese customers to domestic alternatives like Huawei. While Huawei’s latest chips are purportedly approaching Nvidia’s in raw computing power, analysts point out that Huawei still lags in critical areas like software optimization and memory bandwidth.
Adding to the pressure, Chinese state media has raised security concerns surrounding Nvidia’s chips, with regulators cautioning domestic firms against purchasing the H20, a claim Nvidia vehemently denies.
Another chip in the works
In addition to the B30A, Nvidia is also reportedly developing another Blackwell-based chip specifically designed for AI inference tasks in China. This RTX6000D chip is expected to be more affordable than the H20 thanks to a simpler design and reduced specifications.
Reuters previously reported that the RTX6000D is engineered to operate just below the performance thresholds outlined by US export restrictions. Utilizing standard GDDR memory and operating at 1,398 gigabytes per second, it falls slightly below the 1.4 terabyte-per-second limit imposed by the new regulations in April.
Initial shipments of the RTX6000D are anticipated to reach Chinese customers as early as September, signaling Nvidia’s continued efforts to navigate the complex regulatory landscape and serve the Chinese market.
(Photo by BoliviaInteligente)
“`
Original article, Author: Samuel Thompson. If you wish to reprint this article, please indicate the source:https://aicnbc.com/7664.html